perf: add benchmark, optimize

Realized about a 10x speedup.

- add benches/million-updates.rs
- optimization: use rand instead of nanorand (because of thread_rng() initialization being muchf aster)
- update README

Signed-off-by: Stephen Waits <steve@waits.net>

.gitignore

@@ -1,2 +1,3 @@
+/flamegraph.svg
 /target
 Cargo.lock

Cargo.toml

@@ -20,13 +20,24 @@ default = ["f32"]
 f32 = []
 f64 = []
 
+[profile.bench]
+debug = true
+
+[profile.release]
+debug = true
+
 [dev-dependencies]
 assert_approx_eq = "1.1.0"
+criterion = "0.4.0"
 proptest = "1.0.0"
 serde_json = "1.0.91"
 
+[[bench]]
+name = "million-updates"
+harness = false
+
 [dependencies]
-nanorand = {version = "0.7.0", features = ["wyrand", "getrandom"]}
 num-traits = "0.2.15"
+rand = "0.8.5"
 serde = {version = "1.0.152", features = ["serde_derive"]}
 serde_with = "2.2.0"

README.md

@@ -172,10 +172,11 @@ One design goal of this crate is to minimize dependencies. When dependencies
 are required, I try to be very selective about them (re: bloat and licensing).
 This crate depends on:
 
-- [`nanorand`](https://crates.io/crates/nanorand): A fast PRNG based on WyRand.
+- [`rand`](https://crates.io/crates/rand): I started with `nanorand`; however,
-  It pulls in `getrandom` as a portable source of entropy. I chose this over
+  this is faster because of `rand::thread_rng()`; whereas with `nanorand` we
-  the more commonly used, but heavyweight `rand` because it's significantly
+  ended up needing to reimplement that same functionality through something
-  smaller and does just what I need, and not much more.
+  like `lazy_static!` plus `Mutex`, which is not worth it for a few percentage
+  points of performance.
 - [`num-traits`](https://crates.io/crates/num-traits): A brilliant crate which
   makes it easier to operate on numbers in generics; like using `Float` as a
   constraint on a generic type. Its convenience outweighs its cost.
@@ -190,7 +191,8 @@ This crate depends on:
 
 Several crates implemented Vivaldi NC before this one. So, why another?
 
-I had several design goals which the existing crates didn't satisfy:
+I had several design goals which the existing crates didn't satisfy, in order
+of priority:
 
 1. Provide the *simplest interface possible*. I just want to have some sort of
    Coordinate struct and be able to update it, and then use it to estimate
@@ -202,7 +204,10 @@ I had several design goals which the existing crates didn't satisfy:
 3. Serializable/Deserializable by default. These things are useful for sending
    across networks. The intent is to be able to do that without a bunch of
    rigamarole. Support `serde` traits by default.
-4. Well documented. Well tested. This varies by the crates.
+4. Well documented. Well tested. This varies in the existing crates.
+5. Performance. While it's a low priority, it should be reasonable. The library
+   currently does about 10,000 `NetworkCoordinate::update()` calls per
+   millisecond on my Macbook Pro (Intel i7-9750H (12) @ 2.60GHz).
 
 ### Other Vivaldi NC Implementations
 

benches/million-updates.rs

@@ -0,0 +1,52 @@
+use std::time::Duration;
+
+use criterion::{criterion_group, criterion_main, Criterion};
+use vivaldi_nc::{NetworkCoordinate2D, NetworkCoordinate3D};
+
+const NUM_NODES: usize = 1_000;
+
+fn million_vivaldi_2d_updates() {
+    // create NUM_NODES NCs
+    let mut nc: Vec<NetworkCoordinate2D> =
+        (0..NUM_NODES).map(|_| NetworkCoordinate2D::new()).collect();
+
+    // loop, update each item for every other item
+    for i in 0..NUM_NODES {
+        let nc_i = nc[i].clone();
+        for (j, nc_j) in nc.iter_mut().enumerate().take(NUM_NODES) {
+            if i == j {
+                continue;
+            }
+            let rtt = if i > j { i - j } else { j - i };
+            // let nc_j = nc[j].clone();
+            nc_j.update(&nc_i, Duration::from_millis(rtt.try_into().unwrap()));
+        }
+    }
+}
+
+fn million_vivaldi_3d_updates() {
+    // create NUM_NODES NCs
+    let mut nc: Vec<NetworkCoordinate3D> =
+        (0..NUM_NODES).map(|_| NetworkCoordinate3D::new()).collect();
+
+    // loop, update each item for every other item
+    for i in 0..NUM_NODES {
+        let nc_i = nc[i].clone();
+        for (j, nc_j) in nc.iter_mut().enumerate().take(NUM_NODES) {
+            if i == j {
+                continue;
+            }
+            let rtt = if i > j { i - j } else { j - i };
+            // let nc_j = nc[j].clone();
+            nc_j.update(&nc_i, Duration::from_millis(rtt.try_into().unwrap()));
+        }
+    }
+}
+
+fn run_benchmarks(c: &mut Criterion) {
+    c.bench_function("million 2D updates", |b| b.iter(million_vivaldi_2d_updates));
+    c.bench_function("million 3D updates", |b| b.iter(million_vivaldi_3d_updates));
+}
+
+criterion_group!(benches, run_benchmarks);
+criterion_main!(benches);

examples/planetlab.rs

@@ -8,7 +8,7 @@ use std::{
     time::Duration,
 };
 
-use nanorand::{Rng, WyRand};
+use rand::Rng;
 use vivaldi_nc::NetworkCoordinate;
 
 // this will return a Vec<Vec<Duration>>
@@ -66,10 +66,10 @@ fn main() {
     // 4. adjust local NC
     let mut error = 0.0;
     for _ in 0..30_000 {
-        let mut rng = WyRand::new();
+        let mut rng = rand::thread_rng();
         // 1. choose a local and remote NC
-        let i_local = rng.generate::<usize>() % NUM_NODES;
+        let i_local = rng.gen_range(0..NUM_NODES);
-        let i_remote = rng.generate::<usize>() % NUM_NODES;
+        let i_remote = rng.gen_range(0..NUM_NODES);
         if i_local == i_remote {
             continue;
         }

proptest-regressions/height_vector.txt

@@ -9,3 +9,4 @@ cc 38ed1e2c313317809f5469d856d2c1c61e4851c64a156cd1bdd0757fa7c77128 # shrinks to
 cc b7d1b5e9fe3e33066b583c4532865373179972e9b5786ecc62438f23bc939e87 # shrinks to x = -0.0, y = 0.0, h = -3.448118e-33, m = -0.0
 cc 3c4fc4f21c82ffe5bcbc740ca9ab598e0894226acb86a3046a3a24cbbcdbf8e6 # shrinks to x = 0.0, y = 0.0, h = -0.0
 cc d1a551d7eb92bf69e3d0cd0b677baac5aaa5a95d6a016fccf67cf57af71c8b0c # shrinks to x = -5.098216e26, y = 0.0, h = 0.0
+cc 7fa6f2f56f2398bae6804c5f0c0f3f47a7304e505c34f3addb676546cefd97f2 # shrinks to x = -0.0, y = 1.500389e38, h = 2.828499e38

src/height_vector.rs

@@ -44,7 +44,7 @@
 
 use core::ops::{Add, Mul, Sub};
 
-use nanorand::{Rng, WyRand};
+use rand::prelude::*;
 use serde::{Deserialize, Serialize};
 
 use crate::vector::{self, Vector};
@@ -99,12 +99,12 @@ impl<const N: usize> HeightVector<N> {
 
     /// A new height vector is a random unit vector
     pub(crate) fn random() -> Self {
-        let mut rng = WyRand::new();
+        let mut rng = rand::thread_rng();
         let mut vec = [0.0; N];
         for i in vec.iter_mut().take(N) {
-            *i = rng.generate::<FloatType>();
+            *i = rng.gen::<FloatType>() - 0.5;
         }
-        let height = rng.generate::<FloatType>().abs();
+        let height = rng.gen::<FloatType>().abs();
         Self {
             position: Vector::<FloatType, N>::from(vec),
             height,
@@ -255,7 +255,11 @@ mod tests {
                 assert_approx_eq!(a.len(), 1.0);
             } else {
                 // `HeightVector` we created from proptest values should be valid and have a length
-                assert_approx_eq!(a.len(), len);
+                if len.is_nan() || len.is_infinite() {
+                    assert!(a.len().is_nan() || a.len().is_infinite());
+                } else {
+                    assert_approx_eq!(a.len(), len);
+                }
             }
         }